arch/sparc64/mm/hugetlbpage.c

   1 /*
   2  * SPARC64 Huge TLB page support.
   3  *
   4  * Copyright (C) 2002, 2003, 2006 David S. Miller (davem@davemloft.net)
   5  */
   6
   7 #include <linux/init.h>
   8 #include <linux/module.h>
   9 #include <linux/fs.h>
  10 #include <linux/mm.h>
  11 #include <linux/hugetlb.h>
  12 #include <linux/pagemap.h>
  13 #include <linux/smp_lock.h>
  14 #include <linux/slab.h>
  15 #include <linux/sysctl.h>
  16
  17 #include <asm/mman.h>
  18 #include <asm/pgalloc.h>
  19 #include <asm/tlb.h>
  20 #include <asm/tlbflush.h>
  21 #include <asm/cacheflush.h>
  22 #include <asm/mmu_context.h>
  23
  24 /* Slightly simplified from the non-hugepage variant because by
  25  * definition we don't have to worry about any page coloring stuff
  26  */
  27 #define VA_EXCLUDE_START (0x0000080000000000UL - (1UL << 32UL))
  28 #define VA_EXCLUDE_END   (0xfffff80000000000UL + (1UL << 32UL))
  29
  30 static unsigned long hugetlb_get_unmapped_area_bottomup(struct file *filp,
  31                                                         unsigned long addr,
  32                                                         unsigned long len,
  33                                                         unsigned long pgoff,
  34                                                         unsigned long flags)
  35 {
  36         struct mm_struct *mm = current->mm;
  37         struct vm_area_struct * vma;
  38         unsigned long task_size = TASK_SIZE;
  39         unsigned long start_addr;
  40
  41         if (test_thread_flag(TIF_32BIT))
  42                 task_size = STACK_TOP32;
  43         if (unlikely(len >= VA_EXCLUDE_START))
  44                 return -ENOMEM;
  45
  46         if (len > mm->cached_hole_size) {
  47                 start_addr = addr = mm->free_area_cache;
  48         } else {
  49                 start_addr = addr = TASK_UNMAPPED_BASE;
  50                 mm->cached_hole_size = 0;
  51         }
  52
  53         task_size -= len;
  54
  55 full_search:
  56         addr = ALIGN(addr, HPAGE_SIZE);
  57
  58         for (vma = find_vma(mm, addr); ; vma = vma->vm_next) {
  59                 /* At this point:  (!vma || addr < vma->vm_end). */
  60                 if (addr < VA_EXCLUDE_START &&
  61                     (addr + len) >= VA_EXCLUDE_START) {
  62                         addr = VA_EXCLUDE_END;
  63                         vma = find_vma(mm, VA_EXCLUDE_END);
  64                 }
  65                 if (unlikely(task_size < addr)) {
  66                         if (start_addr != TASK_UNMAPPED_BASE) {
  67                                 start_addr = addr = TASK_UNMAPPED_BASE;
  68                                 mm->cached_hole_size = 0;
  69                                 goto full_search;
  70                         }
  71                         return -ENOMEM;
  72                 }
  73                 if (likely(!vma || addr + len <= vma->vm_start)) {
  74                         /*
  75                          * Remember the place where we stopped the search:
  76                          */
  77                         mm->free_area_cache = addr + len;
  78                         return addr;
  79                 }
  80                 if (addr + mm->cached_hole_size < vma->vm_start)
  81                         mm->cached_hole_size = vma->vm_start - addr;
  82
  83                 addr = ALIGN(vma->vm_end, HPAGE_SIZE);
  84         }
  85 }
  86
  87 static unsigned long
  88 hugetlb_get_unmapped_area_topdown(struct file *filp, const unsigned long addr0,
  89                                   const unsigned long len,
  90                                   const unsigned long pgoff,
  91                                   const unsigned long flags)
  92 {
  93         struct vm_area_struct *vma;
  94         struct mm_struct *mm = current->mm;
  95         unsigned long addr = addr0;
  96
  97         /* This should only ever run for 32-bit processes.  */
  98         BUG_ON(!test_thread_flag(TIF_32BIT));
  99
 100         /* check if free_area_cache is useful for us */
 101         if (len <= mm->cached_hole_size) {
 102                 mm->cached_hole_size = 0;
 103                 mm->free_area_cache = mm->mmap_base;
 104         }
 105
 106         /* either no address requested or can't fit in requested address hole */
 107         addr = mm->free_area_cache & HPAGE_MASK;
 108
 109         /* make sure it can fit in the remaining address space */
 110         if (likely(addr > len)) {
 111                 vma = find_vma(mm, addr-len);
 112                 if (!vma || addr <= vma->vm_start) {
 113                         /* remember the address as a hint for next time */
 114                         return (mm->free_area_cache = addr-len);
 115                 }
 116         }
 117
 118         if (unlikely(mm->mmap_base < len))
 119                 goto bottomup;
 120
 121         addr = (mm->mmap_base-len) & HPAGE_MASK;
 122
 123         do {
 124                 /*
 125                  * Lookup failure means no vma is above this address,
 126                  * else if new region fits below vma->vm_start,
 127                  * return with success:
 128                  */
 129                 vma = find_vma(mm, addr);
 130                 if (likely(!vma || addr+len <= vma->vm_start)) {
 131                         /* remember the address as a hint for next time */
 132                         return (mm->free_area_cache = addr);
 133                 }
 134
 135                 /* remember the largest hole we saw so far */
 136                 if (addr + mm->cached_hole_size < vma->vm_start)
 137                         mm->cached_hole_size = vma->vm_start - addr;
 138
 139                 /* try just below the current vma->vm_start */
 140                 addr = (vma->vm_start-len) & HPAGE_MASK;
 141         } while (likely(len < vma->vm_start));
 142
 143 bottomup:
 144         /*
 145          * A failed mmap() very likely causes application failure,
 146          * so fall back to the bottom-up function here. This scenario
 147          * can happen with large stack limits and large mmap()
 148          * allocations.
 149          */
 150         mm->cached_hole_size = ~0UL;
 151         mm->free_area_cache = TASK_UNMAPPED_BASE;
 152         addr = arch_get_unmapped_area(filp, addr0, len, pgoff, flags);
 153         /*
 154          * Restore the topdown base:
 155          */
 156         mm->free_area_cache = mm->mmap_base;
 157         mm->cached_hole_size = ~0UL;
 158
 159         return addr;
 160 }
 161
 162 unsigned long
 163 hugetlb_get_unmapped_area(struct file *file, unsigned long addr,
 164                 unsigned long len, unsigned long pgoff, unsigned long flags)
 165 {
 166         struct mm_struct *mm = current->mm;
 167         struct vm_area_struct *vma;
 168         unsigned long task_size = TASK_SIZE;
 169
 170         if (test_thread_flag(TIF_32BIT))
 171                 task_size = STACK_TOP32;
 172
 173         if (len & ~HPAGE_MASK)
 174                 return -EINVAL;
 175         if (len > task_size)
 176                 return -ENOMEM;
 177
 178         if (addr) {
 179                 addr = ALIGN(addr, HPAGE_SIZE);
 180                 vma = find_vma(mm, addr);
 181                 if (task_size - len >= addr &&
 182                     (!vma || addr + len <= vma->vm_start))
 183                         return addr;
 184         }
 185         if (mm->get_unmapped_area == arch_get_unmapped_area)
 186                 return hugetlb_get_unmapped_area_bottomup(file, addr, len,
 187                                 pgoff, flags);
 188         else
 189                 return hugetlb_get_unmapped_area_topdown(file, addr, len,
 190                                 pgoff, flags);
 191 }
 192
 193 pte_t *huge_pte_alloc(struct mm_struct *mm, unsigned long addr)
 194 {
 195         pgd_t *pgd;
 196         pud_t *pud;
 197         pmd_t *pmd;
 198         pte_t *pte = NULL;
 199
 200         /* We must align the address, because our caller will run
 201          * set_huge_pte_at() on whatever we return, which writes out
 202          * all of the sub-ptes for the hugepage range.  So we have
 203          * to give it the first such sub-pte.
 204          */
 205         addr &= HPAGE_MASK;
 206
 207         pgd = pgd_offset(mm, addr);
 208         pud = pud_alloc(mm, pgd, addr);
 209         if (pud) {
 210                 pmd = pmd_alloc(mm, pud, addr);
 211                 if (pmd)
 212                         pte = pte_alloc_map(mm, pmd, addr);
 213         }
 214         return pte;
 215 }
 216
 217 pte_t *huge_pte_offset(struct mm_struct *mm, unsigned long addr)
 218 {
 219         pgd_t *pgd;
 220         pud_t *pud;
 221         pmd_t *pmd;
 222         pte_t *pte = NULL;
 223
 224         addr &= HPAGE_MASK;
 225
 226         pgd = pgd_offset(mm, addr);
 227         if (!pgd_none(*pgd)) {
 228                 pud = pud_offset(pgd, addr);
 229                 if (!pud_none(*pud)) {
 230                         pmd = pmd_offset(pud, addr);
 231                         if (!pmd_none(*pmd))
 232                                 pte = pte_offset_map(pmd, addr);
 233                 }
 234         }
 235         return pte;
 236 }
 237
 238 void set_huge_pte_at(struct mm_struct *mm, unsigned long addr,
 239                      pte_t *ptep, pte_t entry)
 240 {
 241         int i;
 242
 243         if (!pte_present(*ptep) && pte_present(entry))
 244                 mm->context.huge_pte_count++;
 245
 246         for (i = 0; i < (1 << HUGETLB_PAGE_ORDER); i++) {
 247                 set_pte_at(mm, addr, ptep, entry);
 248                 ptep++;
 249                 addr += PAGE_SIZE;
 250                 pte_val(entry) += PAGE_SIZE;
 251         }
 252 }
 253
 254 pte_t huge_ptep_get_and_clear(struct mm_struct *mm, unsigned long addr,
 255                               pte_t *ptep)
 256 {
 257         pte_t entry;
 258         int i;
 259
 260         entry = *ptep;
 261         if (pte_present(entry))
 262                 mm->context.huge_pte_count--;
 263
 264         for (i = 0; i < (1 << HUGETLB_PAGE_ORDER); i++) {
 265                 pte_clear(mm, addr, ptep);
 266                 addr += PAGE_SIZE;
 267                 ptep++;
 268         }
 269
 270         return entry;
 271 }
 272
 273 struct page *follow_huge_addr(struct mm_struct *mm,
 274                               unsigned long address, int write)
 275 {
 276         return ERR_PTR(-EINVAL);
 277 }
 278
 279 int pmd_huge(pmd_t pmd)
 280 {
 281         return 0;
 282 }
 283
 284 struct page *follow_huge_pmd(struct mm_struct *mm, unsigned long address,
 285                              pmd_t *pmd, int write)
 286 {
 287         return NULL;
 288 }
 289
 290 static void context_reload(void *__data)
 291 {
 292         struct mm_struct *mm = __data;
 293
 294         if (mm == current->mm)
 295                 load_secondary_context(mm);
 296 }
 297
 298 void hugetlb_prefault_arch_hook(struct mm_struct *mm)
 299 {
 300         struct tsb_config *tp = &mm->context.tsb_block[MM_TSB_HUGE];
 301
 302         if (likely(tp->tsb != NULL))
 303                 return;
 304
 305         tsb_grow(mm, MM_TSB_HUGE, 0);
 306         tsb_context_switch(mm);
 307         smp_tsb_sync(mm);
 308
 309         /* On UltraSPARC-III+ and later, configure the second half of
 310          * the Data-TLB for huge pages.
 311          */
 312         if (tlb_type == cheetah_plus) {
 313                 unsigned long ctx;
 314
 315                 spin_lock(&ctx_alloc_lock);
 316                 ctx = mm->context.sparc64_ctx_val;
 317                 ctx &= ~CTX_PGSZ_MASK;
 318                 ctx |= CTX_PGSZ_BASE << CTX_PGSZ0_SHIFT;
 319                 ctx |= CTX_PGSZ_HUGE << CTX_PGSZ1_SHIFT;
 320
 321                 if (ctx != mm->context.sparc64_ctx_val) {
 322                         /* When changing the page size fields, we
 323                          * must perform a context flush so that no
 324                          * stale entries match.  This flush must
 325                          * occur with the original context register
 326                          * settings.
 327                          */
 328                         do_flush_tlb_mm(mm);
 329
 330                         /* Reload the context register of all processors
 331                          * also executing in this address space.
 332                          */
 333                         mm->context.sparc64_ctx_val = ctx;
 334                         on_each_cpu(context_reload, mm, 0, 0);
 335                 }
 336                 spin_unlock(&ctx_alloc_lock);
 337         }
 338 }